What a local AI test actually measured.

A useful measurement starts with the machine and the task.

Review Reply Writer has retained evidence of Gemma generating replies through Ollama. This reference explains one three-case test from October 13, 2026, Hawaii time. Its fictional reviews cover cafe praise, a complaint, and a mixed accommodation experience.

The measurements come from a GitHub-hosted Ubuntu CPU runner. They do not describe my MSI laptop or a test of my Windows setup. Here, local inference means the application called Ollama on the same runner.

Recorded historical execution. The public files contain model responses, timings, settings, and source fingerprints. They contain no human usefulness assessments, manual time baseline, or matched hosted-model comparison. Later code changes make these historical results rather than a current release check.

The configuration behind these numbers.

The project account supplies the server launch settings. The JSON records context as the server default, not independently inspected. Its local_execution_verified field is false: a loopback connection and model metadata do not constitute a network audit.

For a comparable new run, record the actual machine, model digest, settings, source version, and inputs again. Installing the same model name alone does not reproduce these conditions.

For an original Windows workflow rather than these Ubuntu runner results, read the gaming-laptop local AI guide. It includes a loopback-only test script and leaves laptop throughput unmeasured until a real run is documented.

What completed, and what earlier failures taught.

These values convert each JSON elapsed_ms field to seconds. All three drafts passed that version's automated checks without a retry. The complaint preserved the specific wait and ignored question; the mixed reply mentioned both complaints and praise. No person rated the outputs in this record.

The earlier test remains visible. Its checker wrongly rejected wording about what the barista should have done. An accepted retry softened an exact wait into a general delay. The mixed case ended with a labelled template fallback after two rejected model attempts.

Those failures show why a completed response, an automated pass, and a useful reply need separate records. Three selected cases across changed implementations cannot establish a general acceptance rate or a controlled speed improvement. The full Review Reply Writer case explains the product decisions behind the checks.

Nineteen seconds of elapsed processing is not nineteen seconds saved.

The table records case elapsed time around generation and application checks. Model-response receipts have their own durations. Neither measures the person's work preparing a review, checking the draft, correcting it, approving it, or maintaining the setup.

For a whole-task comparison, measure those activities and the usual method at an acceptable common quality. Record waiting separately; it counts as occupied time only when it prevents other useful work. The time calculator explains this accounting. Missing preparation, review, or baseline measurements must remain unknown, rather than being entered as zero.

These records also lack measured hardware, electricity, setup, and human-review costs. They cannot establish a total cost or net saving. There is no matched hosted-model run of these same reviews in the inspected evidence.

Keep the evidence with its limits.

Download the compact factual record (.json), including model identity, exact timings, historical source fingerprints, and unmeasured fields.

• Later raw test record, used for the timing table.
• Earlier raw test record, preserving rejected attempts and fallback.

The links pin the public evidence snapshot at 1c5e0487. That snapshot preserves older results; it is not the tested application version. The reports' individual source fingerprints identify the historical code, and the documentation explicitly notes subsequent changes.

Prepared with AI assistance from the linked public records. No new model run, laptop access, paid inference, or human evaluation was performed for this reference.

Explore practical AI · Send a correction